0G Compute Inference

0G Compute Network provides decentralized AI inference services, supporting various AI models including Large Language Models (LLM), text-to-image generation, and speech-to-text processing.

Prerequisites

• Node.js >= 22.0.0
• A wallet with 0G tokens (either testnet or mainnet)
• EVM compatible wallet (for Web UI)

Supported Service Types

• Chatbot Services: Conversational AI with models like GPT, DeepSeek, and others
• Text-to-Image: Generate images from text descriptions using Stable Diffusion and similar models
• Speech-to-Text: Transcribe audio to text using Whisper and other speech recognition models

Available Services

Testnet Services

View Testnet Services (3 Available)

#	Model	Type	Provider	URL	Input Price (per token)	Output Price (per token)
1	qwen/qwen-2.5-7b-instruct	Chatbot	0xa48f01...	compute-network-6	0.00000005 0G	0.0000001 0G
2	openai/gpt-oss-20b	Chatbot	0x8e60d4...	compute-network-7	0.00000005 0G	0.0000001 0G
3	google/gemma-3-27b-it	Chatbot	0x69Eb5a...	compute-network-8	0.00000015 0G	0.0000004 0G

Available Models:

• Qwen 2.5 7B Instruct: Fast and efficient conversational model
• GPT-OSS-20B: Mid-size open-source GPT alternative
• Gemma 3 27B IT: Google's instruction-tuned model

All testnet services feature TeeML verifiability and are ideal for development and testing.

Mainnet Services

View Mainnet Services (7 Available)

#	Model	Type	Provider	Input Price (per token)	Output Price (per token)
1	deepseek-ai/DeepSeek-V3.1	Chatbot	0xd9966e...	0.00000049 0G	0.00000015 0G
2	openai/whisper-large-v3	Speech-to-Text	0x36aCff...	0.000000049 0G	0.000000114 0G
3	openai/gpt-oss-120b	Chatbot	0xBB3f5b...	0.0000001 0G	0.00000049 0G
4	qwen/qwen2.5-vl-72b-instruct	Chatbot	0x4415ef...	0.00000049 0G	0.00000049 0G
5	deepseek/deepseek-chat-v3-0324	Chatbot	0x1B3AAe...	0.0000003 0G	0.000001 0G
6	flux-turbo	Text-to-Image	0xE29a72...	0.0 0G	0.003 0G
7	openai/gpt-oss-20b	Chatbot	0x44ba50...	0.00000005 0G	0.00000011 0G

Available Models by Type:

Chatbots (5 models):

• DeepSeek V3.1: Latest high-performance reasoning model
• GPT-OSS-120B: Large-scale open-source GPT model
• Qwen 2.5 VL 72B: Vision-language multimodal model
• DeepSeek Chat V3: Optimized conversational model
• GPT-OSS-20B: Efficient mid-size model

Speech-to-Text (1 model):

• Whisper Large V3: OpenAI's state-of-the-art transcription model

Text-to-Image (1 model):

• Flux Turbo: Fast high-quality image generation

All mainnet services feature TeeML verifiability for trusted execution in production environments.

Choose Your Interface

Feature	Web UI	CLI	SDK
Setup time	~1 min	~2 min	~5 min
Interactive chat	✅	❌	❌
Automation	❌	✅	✅
App integration	❌	❌	✅
Direct API access	❌	❌	✅

Web UI

Best for: Quick testing, experimentation and direct frontend integration.

Installation

pnpm add @0glabs/0g-serving-broker -g

Launch Web UI

0g-compute-cli ui start-web

Open http://localhost:3090 in your browser.

Getting Started

1. Connect & Fund

1. Connect your wallet (MetaMask recommended)
2. Deposit some 0G tokens using the account dashboard
3. Browse available AI models and their pricing

2. Start Using AI Services

Option A: Chat Interface

• Click "Chat" on any chatbot provider
• Start conversations immediately
• Perfect for testing and experimentation

Option B: Get API Integration

• Click "Build" on any provider
• Get step-by-step integration guides
• Copy-paste ready code examples

CLI

Best for: Automation, scripting, and server environments

Installation

pnpm add @0glabs/0g-serving-broker -g

Setup Environment

Choose Network

0g-compute-cli setup-network

Login with Wallet

Enter your wallet private key when prompted. This will be used for account management and service payments.

0g-compute-cli login

Create Account & Add Funds

Before using inference services, you need to fund your account. For detailed account management, see Account Management.

0g-compute-cli deposit --amount 10
0g-compute-cli get-account
0g-compute-cli transfer-fund --provider <PROVIDER_ADDRESS> --amount 5

CLI Commands

List Providers

0g-compute-cli inference list-providers

Verify Provider

Check provider's TEE attestation and reliability before using:

0g-compute-cli inference verify --provider <PROVIDER_ADDRESS>

This command outputs the provider's report and verifies their Trusted Execution Environment (TEE) status.

Acknowledge Provider

Before using a provider, acknowledge them on-chain:

0g-compute-cli inference acknowledge-provider --provider <PROVIDER_ADDRESS>

Direct API Access

Generate an authentication token for direct API calls:

0g-compute-cli inference get-secret --provider <PROVIDER_ADDRESS>

This generates a Bearer token in the format app-sk-<SECRET> that you can use for direct API calls.

API Usage Examples

Chatbot

Use for conversational AI and text generation.

cURL

curl <service_url>/v1/proxy/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer app-sk-<YOUR_SECRET>" \
  -d '{
    "model": <service.model>,
    "messages": [
      {
        "role": "system",
        "content": "You are a helpful assistant."
      },
      {
        "role": "user",
        "content": "Hello!"
      }
    ]
  }`

JavaScript

const OpenAI = require('openai');

const client = new OpenAI({
  baseURL: `${service.url}/v1/proxy`,
  apiKey: 'app-sk-<YOUR_SECRET>'
});

const completion = await client.chat.completions.create({
  model: service.model,
  messages: [
    {
      role: 'system',
      content: 'You are a helpful assistant.'
    },
    {
      role: 'user',
      content: 'Hello!'
    }
  ]
});

console.log(completion.choices[0].message);

Python

from openai import OpenAI

client = OpenAI(
    base_url=`${service.url}/v1/proxy`,
    api_key='app-sk-<YOUR_SECRET>'
)

completion = client.chat.completions.create(
    model=service.model,
    messages=[
        {
            'role': 'system',
            'content': 'You are a helpful assistant.'
        },
        {
            'role': 'user',
            'content': 'Hello!'
        }
    ]
)

print(completion.choices[0].message)

Text-to-Image

Generate images from text descriptions.

cURL

curl <service_url>/v1/proxy/images/generations \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer app-sk-<YOUR_SECRET>" \
  -d '{
    "model": <service.model>,
    "prompt": "A cute baby sea otter playing in the water",
    "n": 1,
    "size": "1024x1024"
  }'

JavaScript

const OpenAI = require('openai');

const client = new OpenAI({
  baseURL: `${service.url}/v1/proxy`,
  apiKey: 'app-sk-<YOUR_SECRET>'
});

const response = await client.images.generate({
  model: service.model,
  prompt: 'A cute baby sea otter playing in the water',
  n: 1,
  size: '1024x1024'
});

console.log(response.data);

Python

from openai import OpenAI

client = OpenAI(
    base_url=`${service.url}/v1/proxy`,
    api_key='app-sk-<YOUR_SECRET>'
)

response = client.images.generate(
    model=service.model,
    prompt='A cute baby sea otter playing in the water',
    n=1,
    size='1024x1024'
)

print(response.data)

Speech-to-Text

Transcribe audio files to text.

cURL

curl <service_url>/v1/proxy/audio/transcriptions \
  -H "Authorization: Bearer app-sk-<YOUR_SECRET>" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@audio.ogg" \
  -F "model=whisper-large-v3" \
  -F "response_format=json"

JavaScript

const OpenAI = require('openai');
const fs = require('fs');

const client = new OpenAI({
  baseURL: `${service.url}/v1/proxy`,
  apiKey: 'app-sk-<YOUR_SECRET>'
});

const transcription = await client.audio.transcriptions.create({
  file: fs.createReadStream('audio.ogg'),
  model: 'whisper-large-v3',
  response_format: 'json'
});

console.log(transcription.text);

Python

from openai import OpenAI

client = OpenAI(
    base_url=`${service.url}/v1/proxy`,
    api_key='app-sk-<YOUR_SECRET>'
)

with open('audio.ogg', 'rb') as audio_file:
    transcription = client.audio.transcriptions.create(
        file=audio_file,
        model='whisper-large-v3',
        response_format='json'
    )

print(transcription.text)

Start Local Proxy Server

Run a local OpenAI-compatible server:

# Start server on port 3000 (default)
0g-compute-cli inference serve --provider <PROVIDER_ADDRESS>

# Custom port
0g-compute-cli inference serve --provider <PROVIDER_ADDRESS> --port 8080

Then use any OpenAI-compatible client to connect to http://localhost:3000.

SDK

Best for: Application integration and programmatic access

Installation

pnpm add @0glabs/0g-serving-broker

Starter Kits Available

Get up and running quickly with our comprehensive TypeScript starter kit within minutes.

• TypeScript Starter Kit - Complete examples with TypeScript and CLI tool

Initialize the Broker

Node.js

import { ethers } from "ethers";
import { createZGComputeNetworkBroker } from "@0glabs/0g-serving-broker";

// Choose your network
const RPC_URL = process.env.NODE_ENV === 'production'
  ? "https://evmrpc.0g.ai"  // Mainnet
  : "https://evmrpc-testnet.0g.ai";  // Testnet

const provider = new ethers.JsonRpcProvider(RPC_URL);
const wallet = new ethers.Wallet(process.env.PRIVATE_KEY!, provider);
const broker = await createZGComputeNetworkBroker(wallet);

Browser

import { BrowserProvider } from "ethers";
import { createZGComputeNetworkBroker } from "@0glabs/0g-serving-broker";

// Check if MetaMask is installed
if (typeof window.ethereum === "undefined") {
  throw new Error("Please install MetaMask");
}

const provider = new BrowserProvider(window.ethereum);
const signer = await provider.getSigner();
const broker = await createZGComputeNetworkBroker(signer);

Browser Compatibility

@0glabs/0g-serving-broker requires polyfills for Node.js built-in modules.

Vite example:

pnpm add -D vite-plugin-node-polyfills

// vite.config.js
import { nodePolyfills } from 'vite-plugin-node-polyfills';

export default {
  plugins: [
    nodePolyfills({
      include: ['crypto', 'stream', 'util', 'buffer', 'process'],
      globals: { Buffer: true, global: true, process: true }
    })
  ]
};

Discover Services

// List all available services
const services = await broker.inference.listService();

// Filter by service type
const chatbotServices = services.filter(s => s.serviceType === 'chatbot');
const imageServices = services.filter(s => s.serviceType === 'text-to-image');
const speechServices = services.filter(s => s.serviceType === 'speech-to-text');

Account Management

For detailed account operations, see Account Management.

const account = await broker.ledger.getLedger();
await broker.ledger.depositFund(10);
// Required before first use of a provider
await broker.inference.acknowledgeProviderSigner(providerAddress);

Make Inference Requests

Chatbot

const messages = [{ role: "user", content: "Hello!" }];

// Get service metadata
const { endpoint, model } = await broker.inference.getServiceMetadata(providerAddress);

// Generate auth headers
const headers = await broker.inference.getRequestHeaders(
  providerAddress
);

// Make request
const response = await fetch(`${endpoint}/chat/completions`, {
  method: "POST",
  headers: { "Content-Type": "application/json", ...headers },
  body: JSON.stringify({ messages, model })
});

const data = await response.json();
const answer = data.choices[0].message.content;

Text-to-Image

const prompt = "A cute baby sea otter";

const body = JSON.stringify({
    model,
    prompt,
    n: 1,
    size: "1024x1024"
  });

// Get service metadata
const { endpoint, model } = await broker.inference.getServiceMetadata(providerAddress);

// Generate auth headers
const headers = await broker.inference.getRequestHeaders(
  providerAddress,
  body
);

// Make request
const response = await fetch(`${endpoint}/images/generations`, {
  method: "POST",
  headers: { "Content-Type": "application/json", ...headers },
  body: JSON.stringify({
    model,
    prompt,
    n: 1,
    size: "1024x1024"
  })
});

const data = await response.json();
const imageUrl = data.data[0].url;

Speech-to-Text

const formData = new FormData();
formData.append('file', audioFile); // audioFile is a File or Blob
formData.append('model', model);
formData.append('response_format', 'json');

// Get service metadata
const { endpoint, model } = await broker.inference.getServiceMetadata(providerAddress);

// Generate auth headers
const headers = await broker.inference.getRequestHeaders(
  providerAddress
);

// Make request
const response = await fetch(`${endpoint}/audio/transcriptions`, {
  method: "POST",
  headers: { ...headers },
  body: formData
});

const data = await response.json();
const transcription = data.text;

Response Verification

The processResponse method handles response verification and automatic fee management. Both parameters are optional:

• receivedContent: The usage data from the service response. When provided, the SDK caches accumulated usage and automatically transfers funds from your main account to the provider's sub-account to prevent service interruptions.
• chatID: Response identifier for verifiable TEE services. Different service types handle this differently.

Chatbot

For chatbot services, pass the usage data from the response to enable automatic fee management:

// Standard chat completion
const response = await fetch(`${endpoint}/chat/completions`, {
  method: "POST",
  headers: { "Content-Type": "application/json", ...headers },
  body: JSON.stringify({ messages, model })
});

const data = await response.json();

// Process response for automatic fee management
if (data.usage) {
  await broker.inference.processResponse(
    providerAddress,
    undefined,              // chatID is undefined for non-verifiable responses
    JSON.stringify(data.usage)  // Pass usage data for fee calculation
  );
}

// For verifiable TEE services with chatID
// First try to get chatID from response body
let chatID = data.id || data.chatID;

// If not found in response body, check response headers
if (!chatID) {
  chatID = response.headers.get("ZG-Res-Key") || response.headers.get("zg-res-key");
}

if (chatID) {
  const isValid = await broker.inference.processResponse(
    providerAddress,
    chatID,           // Verify the response integrity
    JSON.stringify(data.usage)  // Also manage fees
  );
}

Text-to-Image

For text-to-image services, pass the original request data for input-based fee calculation:

const requestBody = {
  model,
  prompt: "A cute baby sea otter",
  size: "1024x1024",
  n: 1
};

const response = await fetch(`${endpoint}/images/generations`, {
  method: "POST",
  headers: { "Content-Type": "application/json", ...headers },
  body: JSON.stringify(requestBody)
});

const data = await response.json();

// Get chatID from response headers for verification
const chatID = response.headers.get("ZG-Res-Key") || response.headers.get("zg-res-key");

if (chatID) {
  // Process response with chatID for verification and fee calculation
  const isValid = await broker.inference.processResponse(
    providerAddress,
    chatID                        // Verify the response integrity
  );
  console.log("Response valid:", isValid);
} else {
  // Fallback: process without verification if no chatID
  await broker.inference.processResponse(
    providerAddress,
    undefined                     // No chatID available
  );
}

Speech-to-Text

For speech-to-text services, pass the usage data if available:

const formData = new FormData();
formData.append('file', audioFile);
formData.append('model', model);

const response = await fetch(`${endpoint}/audio/transcriptions`, {
  method: "POST",
  headers: { ...headers },
  body: formData
});

const data = await response.json();

// Get chatID from response headers for verification
const chatID = response.headers.get("ZG-Res-Key") || response.headers.get("zg-res-key");

if (chatID) {
  // Process response with chatID for verification and fee calculation
  const isValid = await broker.inference.processResponse(
    providerAddress,
    chatID,                      // Verify the response integrity
    JSON.stringify(data.usage || {}) // Pass usage for fee calculation
  );
  console.log("Response valid:", isValid);
} else if (data.usage) {
  // Fallback: process without verification if no chatID but usage available
  await broker.inference.processResponse(
    providerAddress,
    undefined,                   // No chatID available
    JSON.stringify(data.usage)   // Pass usage for fee calculation
  );
}

Streaming Responses

For streaming responses, handle chatID differently based on service type:

Chatbot Streaming

// For chatbot streaming, first check headers then try to get ID from stream
let chatID = response.headers.get("ZG-Res-Key") || response.headers.get("zg-res-key");

let usage = null;
let streamChatID = null; // Will try to get from stream data
const decoder = new TextDecoder();
const reader = response.body.getReader();

// Process stream
let rawBody = '';
while (true) {
  const { done, value } = await reader.read();
  if (done) break;

  rawBody += decoder.decode(value, { stream: true });
}

// Parse usage and chatID from stream data
for (const line of rawBody.split('\n')) {
  const trimmed = line.trim();
  if (!trimmed || trimmed === 'data: [DONE]') continue;

  try {
    const jsonStr = trimmed.startsWith('data:')
      ? trimmed.slice(5).trim()
      : trimmed;
    const message = JSON.parse(jsonStr);

    // For chatbot, try to get ID from stream data
    if (!streamChatID && (message.id || message.chatID)) {
      streamChatID = message.id || message.chatID;
    }

    if (message.usage) {
      usage = message.usage;
    }
  } catch {}
}

// Use chatID from stream data if available, otherwise use header
const finalChatID = streamChatID || chatID;

// Process with chatID for verification if available
if (finalChatID) {
  const isValid = await broker.inference.processResponse(
    providerAddress,
    finalChatID,
    JSON.stringify(usage || {})
  );
  console.log("Chatbot streaming response valid:", isValid);
} else if (usage) {
  // Fallback: process without verification
  await broker.inference.processResponse(
    providerAddress,
    undefined,
    JSON.stringify(usage)
  );
}

Speech-to-Text Streaming

// For speech-to-text streaming, get chatID from headers
const chatID = response.headers.get("ZG-Res-Key") || response.headers.get("zg-res-key");

let usage = null;
const decoder = new TextDecoder();
const reader = response.body.getReader();

// Process stream
let rawBody = '';
while (true) {
  const { done, value } = await reader.read();
  if (done) break;

  rawBody += decoder.decode(value, { stream: true });
}

// Parse usage from stream data
for (const line of rawBody.split('\n')) {
  const trimmed = line.trim();
  if (!trimmed || trimmed === 'data: [DONE]') continue;

  try {
    const jsonStr = trimmed.startsWith('data:')
      ? trimmed.slice(5).trim()
      : trimmed;
    const message = JSON.parse(jsonStr);
    if (message.usage) {
      usage = message.usage;
    }
  } catch {}
}

// Process with chatID for verification if available
if (chatID) {
  const isValid = await broker.inference.processResponse(
    providerAddress,
    chatID,
    JSON.stringify(usage || {})
  );
  console.log("Audio streaming response valid:", isValid);
} else if (usage) {
  // Fallback: process without verification
  await broker.inference.processResponse(
    providerAddress,
    undefined,
    JSON.stringify(usage)
  );
}

Key Points:

• Always call processResponse after receiving responses to maintain proper fee management
• The SDK automatically handles fund transfers to prevent service interruptions
• For verifiable TEE services, the method also validates response integrity
• chatID retrieval varies by service type:
  • Chatbot: First try response.id or data.chatID, then check ZG-Res-Key header as fallback
  • Text-to-Image & Speech-to-Text: Always get chatID from ZG-Res-Key response header
  • Streaming responses:
    • Chatbot streaming: Check headers first, then try to get id or chatID from stream data
    • Speech-to-text streaming: Get chatID from ZG-Res-Key header immediately
• Usage data format varies by service type but typically includes token counts or request metrics

---

Troubleshooting

Common Issues

Error: Insufficient balance

Your account doesn't have enough funds. Add more using CLI or SDK:

CLI:

Deposit to Main Account

0g-compute-cli deposit --amount 5

Transfer to Provider Sub-Account

0g-compute-cli transfer-fund --provider <PROVIDER_ADDRESS> --amount 5

SDK:

await broker.ledger.depositFund(1);

Error: Provider not acknowledged

You need to acknowledge the provider before using their service:

CLI:

0g-compute-cli inference acknowledge-provider --provider <PROVIDER_ADDRESS>

SDK:

await broker.inference.acknowledgeProviderSigner(providerAddress);

Error: No funds in provider sub-account

Transfer funds to the specific provider sub-account:

0g-compute-cli transfer-fund --provider <PROVIDER_ADDRESS> --amount 5

Check your account balance:

0g-compute-cli get-account

Web UI not starting

If the web UI fails to start:

1. Check if another service is using port 3090:

0g-compute-cli ui start-web --port 3091

2. Ensure the package was installed globally:

pnpm add @0glabs/0g-serving-broker -g

Next Steps

• Manage Accounts → Account Management Guide
• Fine-tune Models → Fine-tuning Guide
• Become a Provider → Provider Setup
• View Examples → GitHub

---

Questions? Join our Discord for support.